Research On Cell Recognition Method For CD56 Image

CD56 is a neural cell adhesion molecule,which can be used in the diagnosis and study on a variety of tumor.CD56 is one of the latest tumor molecular markers,and the study on CD56 images is emerging in the field of digital medical image processing.With the development of deep learning,more and more researchers are applying deep learning techniques to medical image processing in order to assist doctors in diagnosis.Applying deep learning techniques to CD56 images is expected to improve the efficiency and accuracy of cell identification and counting in CD56 images.However,in CD56 images,cells are densely packed,the size of cells varies greatly,the number of positive cells is low compared to negative cells,and there are color differences between cells and between CD56,all of these factors will affect the application of deep learning techniques on CD56 images.To solve the above problems,this paper proposed a series of methods to improve the accuracy of cell recognition in CD56 images based on deep learning,and the specific work can be summarized as the following three parts:(1)We proposed the algorithms of cell annotation refinement and adaptive weighted loss for CD56 images segmentation.Applying the semantic segmentation technique based on deep learning to CD56 images can transform the identification of negative and positive cells into the identification of negative and positive cells pixel points.However,the ratio of the number of pixel points of background,negative cells and positive cells in CD56 images is very unbalanced,which is roughly 70: 10: 1 and will affect the effect of semantic segmentation techniques used in CD56 image segmentation.In our work,we added loss weight to different class of pixels and added adaptive weight to each pixel to improve the loss function of the relevant semantic segmentation model,so that the model could pay more attention to cells,especially positive cells.At the same time,we used clustering method to refine annotations of CD56 images before model training,which could further improve the segmentation accuracy of the model.The experimental results on CD56 image data set show that the refinement of image annotations and the improvement of loss functions in semantic segmentation models used in experiments can effectively improve the segmentation accuracy of models in CD56 images.(2)We proposed the methods of combining Extreme Net with shape constraints and rediscrimination to detect cells from CD56 images.Applying object detection techniques based on deep learning to CD56 images,can identify and detect cells in CD56 images.However,in CD56 images,cells are densely packed and the number of positive cells is low,these affect the effect of object detection techniques used in CD56 images.In our work,we proposed a new hierarchy method to detect cells from CD56 images.Specifically,we first utilize an Extreme Net to roughly detect positive and negative cells from an image.Then shape constraints are adopted to refine the detection results.We further trained a convolutional neural network to identify negative and positive cells more accurately,which is called rediscrimination.Experimental results showed that compared with several state-of-the-art onestage object detection algorithms,the proposed method achieved the highest detection accuracy of cells in CD56 images.(3)We applied the color standardization to CD56 image processing.Due to raw materials and dyes from different suppliers,different image acquisition scanners,different staining protocols of laboratories or hospitals and other reasons,the colors of cells and backgrounds in different images in the same CD56 image data set may be different,which will affect the learning and testing of deep learning model on CD56 images.Thus,the accuracy of CD56 image cell recognition based on deep learning technology will be affected.In our work,we proposed color standardization for CD56 image processing.The color difference between different images in CD56 image data set can be reduced by color standardization.The experimental results show that color standardization can effectively improve the segmentation accuracy of the model in CD56 images.In conclusion,this paper proposed the method for cell annotation refinement and the improved adaptive weighted loss function for CD56 images segmentation.Then based on Extreme Net,this paper proposed the methods of shape constraints and re-discrimination for cell detection in CD56 images.In addition,the color standardization method for CD56 image preprocessing was proposed.We compared the proposed methods with the existing ones and verified the effectiveness of our methods.